Nanbeige4.3-33B Posts 65.5 on DeepSWE Against Qwen3.8-27B's 42.5, With No Independent Run Yet

Nanbeige LLM Lab released Nanbeige4.3-33B, a 33-billion-parameter model under Apache 2.0 with a 524,288-token context, and AI Weekly reported it on October 25. The lab says it leads rivals of similar size on most agentic coding tests, and every figure is self-reported.

The Hugging Face model card shows files open for download with no gating notice, in BF16 Safetensors. It gives no release date. The contact address listed is at kanzhun.com.

A teacher model for a planned 3B

The card describes the model as the flagship teacher for distilling agentic skills into smaller models, such as a planned 3B version. It supports thinking, tool use and multi-turn modes through its chat template.

The card does not describe the training data or method. AI Weekly, written by Alexis Dufresne, quotes the lab's claim of "state-of-the-art agentic performance among open models in the ~30B parameter class".

The numbers against Qwen3.8-27B

The card compares the model with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, mostly at the highest thinking effort. On Terminal-Bench 2.1 it scores 80.9 against 73.0 for the next best. On DeepSWE it scores 65.5 against 42.5.

1. Nanbeige4.3-33B65.5 points
2. Next best (Qwen3.8-27B)42.5 points

Source: Nanbeige4.3-33B model card on Hugging Face, accessed 2026-10-30

Benchmark	Nanbeige4.3-33B	Comparison
Terminal-Bench 2.1	80.9	73.0 next best
DeepSWE	65.5	42.5 next best
Automation Bench	54.2	38.3 next best
AIME 2026	95.2	leads
SWE-Bench Pro	60.3	Qwen3.8-27B 61.7
GDPval	84.8	Qwen3.8-27B 86.4
GPQA Diamond	89.5	Gemma4-31B 91.3

The three rows at the bottom are where it loses. The card says evaluation used specific harnesses and settings. AI Weekly reported the headline claim as wins on 18 of 20 tests, which the card page itself does not total.

What AI Weekly said about the evidence

AI Weekly wrote that "All of these figures are self-reported, with no independent reproduction attached to the release." Its editor added that independent runs will likely shrink the gap before a team commits a stack to the model.

The same piece reports Scicode at 52.6 against 46.7 for Qwen3.8-27B, and AIME 2026 at 95.2 against 94.5. It links the release to Moonshot's Kimi K3.1 Pro, which it says shipped the same day.

Running it

AI Weekly lists example SGLang and vLLM launch commands with tensor-parallel size 4, a Qwen3-coder tool parser and a 262,144-token maximum for reasoning tasks. A preserve_thinking flag keeps reasoning across turns, which the lab recommends for tool-use and agent tasks.

Neither source gives a memory requirement. Nothing there says how the model behaves at the full 524,288-token window.

Where this sits among open weights

Open releases this week differ on terms as well as scores. Mistral's Large 4 restricts hosted inference without a licence, and Nanbeige's Apache 2.0 carries no such clause. DeepSeek's V4.1-Flash-1017 is MIT-licensed and also reports its own benchmark gains.

The 3B distillation the card describes has no date. Independent reproductions of the Terminal-Bench 2.1 and DeepSWE figures are the next thing to watch.